Compatibility of Accelerad with CPU and GPU

[Alex Katsikogiannis]

Dear Accelerad community,

I've been using Radiance, accessed through bifacial_radiance, for light simulation in agri-photovoltaic systems. 

I've primarily used GenDayLit and rtrace for my simulations, which can be quite heavy on computational time for CPU-based tools. For this reason, I decided to upgrade my laptop and utilize Accelerad. 

Most high-end gaming laptops use AMD Ryzen 9 7940HS or 13-generation Intel Core i9-13900HX as the processor. They are mostly equipped with the GeForce RTX 40 series (4070 or 4080) as the graphics card. 

Has anyone used any combination of these hardware for rtrace simulations with Accelerad? Are they compatible with Accelerad?

My background in computer science is quite limited, so any suggestions on hardware and compatibility would be greatly appreciated!

Cheers,

Alex

[Nathaniel Jones]

Hi Alex,

Accelerad is compatible with CUDA-capable (Nvidia) GPUs, which includes your RTX cards.

Nathaniel

[Alex Katsikogiannis]

Hey Nathaniel,

I appreciate the fast response. It is great to hear that Accelerad is compatible with the RTX 40 series. In general, I know that the RTX 40 series have more CUDA cores than the Tesla line; however, does this lead to enhanced performance with Accelerad? 

Furthermore, I am going a bit off-topic here, but I came across this graph in one of your articles:

It seems that there is a minimum number of primary rays, dictated by the -ad Radiance parameter, below which the GPU-based approach is slower than the CPU-based. Is my understanding correct? Furthermore, would this be the case for graphics cards with higher computing capabilities like the RTX 40 series? 

Your feedback is highly appreciated,

Cheers,

Alex

[Nathaniel Jones]

Hi Alex,

I haven't collected data for the RTX 40 series. In general, having more CUDA cores will allow for faster processing for simulations that are large enough to use all of those cores. Other changes in new GPU architectures may also improve simulation times.

Accelerad employs parallelism for primary rays and for irradiance caching, the later of which is related to the -ad parameter. The graph you refer to shows the effect of constant-time operations like copying memory and loading CUDA programs. When the number of parallel operations is small, the simulation time will be dominated by these constant-time operations, in which case it may be faster to use CPU computation. The amount of time taken for these constant-time operations will vary depending on the GPU architecture.

Nathaniel